[developers] predicate naming in MRS

[Woodley Packard]

Hi Stephan,

I’m afraid I have to undermine your scaffold very slightly.  Although ACE preserves case and quoting distinctions internally, users are at least partially shielded from this internal record, in the following ways:

1.  ACE can, in fact, generate quite happily from an MRS with PREDs that are quoted differently from the values found in the ERG.  There appears to be a secondary issue preventing inputs with quoted quantifiers from generating (I try to build an AVM representing that EP, which cannot be made well-formed, because quant_or_wh_relation says PRED is an abstract_q_rel), but a user who didn’t know what convention is used in the ERG could choose to never quote, and expect good results.

2.  The ACE generator manages to ignore case on unquoted PRED values but not for quoted ones.  This is somewhat accidental behavior (resulting from conversion to and from AVMs), but to my point of view, kind of felicituous.  Arguably it should always ignore case -- pending others’ thoughts on the point Ann made about proper names.

3. Similarly, ACE ignores quoting differences when performing MRS comparisons, so _downtown_a_1_rel and "_downtown_a_1_rel" are considered equivalent e.g. in the post-generation subsumption test.  In MRS comparison, case *is* in fact also ignored currently.

I currently fail to be similarly generous about matching transfer rules, although it probably wouldn’t be excessively difficulty to extend the behavior to that domain.

I welcome proposals for improvements to the above behavior. :-)

Merry Christmas and happy new year,
-Woodley


> On Dec 28, 2015, at 9:15 AM, Stephan Oepen <oe at ifi.uio.no> wrote:
> 
> dear colleagues,
> 
> in these days of reflection, i would like to ask for opinions on an
> aspect of the DELPH-IN informalism where dan and i recently discovered
> that we held conflicting opinions.  thus, we are looking for folks
> with a deeper understanding of the issue.
> 
> for MRS predicate symbols, we have long established that we do not
> want case differences or type vs. string distinctions to be
> meaningful, i.e. we do not expect foo, Foo, or "foo" to name different
> relations (see ‘MrsRfc’ on the wiki).  from this, i had concluded that
> no grammar would ever use both foo and "foo", whereas dan has found it
> convenient in the ERG to use comparable type names and strings (across
> lexical entries of different syntactic categories) in the expectation
> that they would be treated as equivalent MRS predicate symbols, e.g.
> _downtown_a_1_rel and "_downtown_a_1_rel".
> 
> my assertation that the above was an undesirable property for a
> DELPH-IN grammar is supported by currrent software: MRS comparison,
> transfer, and generation do not treat types and strings as equivalent;
> a creator of input semantics for generation, for example, needs to
> know about the distinction and make a choice.
> 
> what dan beliefs, however, arguably makes good sense (to me at least).
> i believe i can see how the various pieces of MRS manipulation
> software could be extended to yield the interpretation of equivalence.
> i would volunteer to make these changes in the Lisp implementation of
> MRS-related code.
> 
> before suggesting a course of forward action, i would like to ask (a)
> whether anyone has strongly held positions (and supporting arguments)
> on the general question and (b) whether woodley and mike would be
> prepared to make software changes in ACE and pyDelphin, respectively,
> regarding this choice?
> 
> with thanks in advance, oe
>